基于可能性条件偏好网络的交互式遗传算法及其应用

摘要

根据用户实施的人机交互行为而隐式地获取用户偏好的交互式进化优化算法,可有效减轻用户疲劳,提高个性化搜索或推荐的效率.但是,已有研究没有考虑用户交互行为和偏好的不确定性,影响了对用户偏好的拟合精度以及基于该偏好表达的进化搜索.针对该问题,提出基于可能性条件偏好网络的交互式遗传算法,以刻画用户交互行为和偏好的不确定性,并提高算法的搜索性能.首先,采用交互时间表示交互行为,考虑交互行为的不确定性,给出交互时间可信度的定义,并基于该定义给出了用户不确定偏好的表达函数;其次,利用可信交互时间和偏好函数,定义了用户对评价对象的偏好权重,并利用该权重,设计(更新)可以定量表示用户不确定偏好的可能性条件偏好网络,以更好地拟合用户偏好;然后,结合评价不确定性和可能性条件偏好网络,提出了改进的个体适应值估计策略,以更好地引导搜索;最后,将所提算法应用于图书个性化搜索中,结果表明了算法搜索的可靠性和高效性.%Interactive evolutionary algorithms with user preference implicitly extracted from interactions of user were more powerful in alleviating user fatigue and improving the exploration in personalized search or recom -mendation.However,the uncertainties in user interactions and preferences have not been considered in the previous research,which might greatly impact the reliability of the extracted preference model,as well as the effective exploration of the evolution with that model.Therefore,an interactive genetic algorithm with probabi-listic conditional preference networks(PCP-nets)was proposed,in which,the uncertainties were further fig-ured out according to the interactions,and a PCP-net was designed to depict user preference model with higher accuracy by involving those uncertainties.First, the interaction time was adopted to mathematically describe the relationship between the interactions and user preference,and the reliability of the interaction time was fur-ther defined to reflect the interactive uncertainty.The preference function with evaluation uncertainty was es-tablished with the reliability of interaction time.Second,the preference weights on each interacted object were assigned on the basis of preference function and reliability.With these weights, the PCP-nets were designed and updated by involving the uncertainties into the preference model to improve the approximation.Third, a more accurate fitness function was delivered to assign fitness for the individuals.Last,the proposed algorithm was applied to a personalized book search and its superiority in exploration and feasibility was experimentally demonstrated.